The art and science of hyperparameter optimization on Amazon Nova Forge

This article provides a comprehensive guide to hyperparameter optimization for Amazon Nova Forge, AWS's platform for building custom domain-specific language models. It addresses the critical challenges of fine-tuning while preserving general capabilities.

• Catastrophic forgetting occurs when domain training overwrites general capabilities; data mixing prevents this
• Learning rate is the most sensitive hyperparameter; service defaults are recommended starting points
• Three customization techniques: Continued Pre-Training (CPT), Supervised Fine-Tuning (SFT), Reinforcement Fine-Tuning (RFT)
• Checkpoint selection is the most impactful decision for CPT; match flexibility to data scale
• Data mixing should balance customer data at ~50% with Nova curated data; always include reasoning-instruction-following
• LoRA training offers lower cost and faster iteration; graduate to Full Rank after validation
• Recommended workflow: SFT with LoRA, then RFT for optimal results on labeled data
• Batch size targets: 2-20 million tokens per step for CPT; monitor validation loss for overfitting
• Common mistakes: skipping SFT before RFT, deviating from default learning rates, poor reward functions
• Data and reward quality matter more than volume; prioritize quality over scale

Successful Nova Forge customization requires balancing strategic decisions (checkpoint, data mixing, training mode) with systematic hyperparameter tuning, while maintaining data quality and reward function rigor.